US20030016596A1 - Data transmission from raid devices - Google Patents

Abstract

A redundant array includes a plurality of disks, a bus coupling the disks, a receiving device, and a device to reconstruct a block stored in one of the disks. The device reconstructs the block with associated data and parity blocks from other disks. The device transmits the reconstructed block to the receiving device in response to the one of the disks being unavailable or degraded.

Description

BACKGROUND
• This invention relates generally to the transmission and storage of data on RAID devices and clustered computer systems.[0001]
• Digital video and television systems need high bandwidth data transmission and low latencies. Redundant arrays of inexpensive disks (RAID) support high bandwidth data transfers and very low latencies. RAID storage uses redundancy and/or parity blocks to mask the failure of a disk. RAID arrays have limited capacity to handle failures due to fixed reconstruction cache size.[0002]
• RAID storage divides a received data stream into a sequence of blocks and write consecutive blocks of the sequence to different disks in the array. To retrieve data, the blocks are read from the disks of the array and are reconstituted into the original data stream from the read blocks. To increase reception and transmission speeds, the various disks of the array of the RAID storage may be written to and read from in parallel.[0003]
• Individual disks of a RAID storage configuration will occasionally stall or respond slowly to an access request due to disk surface defects and bad block re-vectoring. Another problem occurs when a drive is being rebuilt or reconstructed from information on other drives. During a slow response, the entire RAID configuration may wait while one disk transmits requested data. Thus, a single slowly responding disk can cause a long latency for a read operation from the RAID configuration.[0004]
• For digital video and cable systems, one slowly responding disk can cause a disaster, because data needs to arrive at a video receiver at a substantially constant rate to keep the receiver's input buffer full. Continued long transmission latencies can deplete the input buffer. If a slow RAID configuration causes a transmission gap so that the receiver's input buffer empties a viewer may perceive a noticeable pause in the video being viewed. Defect-free transmission of video requires that such pauses be absent.[0005]
SUMMARY
• According to an aspect of the invention, a method for reading data from a clustered system of a plurality of redundant array of inexpensive disks (RAID) device configurations includes receiving a message from a requesting one of members of the clustered system that indicates a next location for a particular block for a subsequent read operation from a drive in the RAID device and returning to the requesting one of members of the clustered system a hint that indicates the state of the drive in the RAID device corresponding to location for the subsequent read operation.[0006]
• According to an additional aspect of the invention, a clustered system of a plurality of redundant array of inexpensive disks (RAID) configurations to transmit data blocks to a receiving device includes a plurality of cluster members coupled so that any one cluster member can deliver data to at least some of remaining cluster members. Each cluster member includes a plurality of disks adapted to store the blocks and to transmit the stored blocks to the receiving device and a processor to control reads from and writes to the plurality of disks. The member also includes a process to receive a message that indicates a next location for a particular block for a subsequent read operation from a drive in the RAID device and produce a hint that indicates the state of the drive in the RAID device corresponding to location for the subsequent read operation.[0007]
• According to an additional aspect of the invention, a clustered configuration includes a plurality of clustered nodes each node comprising a redundant array of inexpensive disks (RAID) and a processor adapted to write blocks to and read blocks from the arrays and a process to receive a message that indicates a next location for a particular block for a subsequent read operation from a drive in the redundant arrays of inexpensive disks of one of the nodes and produce a hint that indicates the state of the drive in the redundant arrays of inexpensive disks corresponding to location for the subsequent read operation.[0008]
• One or more of the following advantages may be provided by one or more aspects of the present invention.[0009]
• A clustered system includes a plurality of cluster members. When the clustered system requests data from a RAID configuration of one of the cluster members, a message is return that indicates the state of the member on the next read for a particular block. This message or hint can indicate that reading a particular block will be slow or that the device is degraded or unavailable. When the system reads the particular block, the system will skip the bad drive and instead read the parity and data block from the other drives. A reconstructor performs an XOR operation on parity and data to reconstruct the missing data corresponding to the bad or reconstructing drive. Thus, by skipping the bad drive and not waiting for a slow or degraded node or drive to respond before trying to reconstruct data improves the performance of the RAID array while in degraded state such that performance of the RAID array approaches or reaches normal performance.[0010]
• The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.[0011]
DESCRIPTION OF DRAWINGS
• FIG. 1 is a block diagram of a configuration for a redundant array of inexpensive disks (RAID).[0012]
• FIG. 2 is a block diagram of a clustered computer system employing three redundant array of inexpensive disks (RAID) configurations shown in FIG. 1.[0013]
• FIG. 3 shows a fragment of a data stream sent to the RAID configuration of FIG. 1.[0014]
• FIG. 4 depicts an example of storage of the data fragment of FIG. 3 in the RAID configuration of FIG. 1.[0015]
• FIG. 5 is a block diagram of a reconstructor of data blocks for use in the RAID configuration of FIG. 1.[0016]
• FIG. 6 is a diagram that depicts software layers of a processor used in the RAID configuration of FIG. 1.[0017]
• FIG. 7 is a flow chart depicting a process for reading data from the RAID configuration of FIG. 1.[0018]
• FIG. 8 is a block diagram of a video system using the RAID configuration of FIG. 1..[0019]
DETAILED DESCRIPTION
• Referring to FIG. 1, a[0020]computer node11 includesRAID configuration12 and aprocessor20. TheRAID configuration12 is under control of theprocessor20 and includes threestorage disks14a,14band14cand individual disk controllers (not shown). TheRAID configuration12 also includes abus16 for transmitting data writes to and reads of the threedisks14a,14band14cunder control ofprocessor20. Generally,such configurations12 may have N disks where N is 3 or greater. In some embodiments two disk drives can be used if mirroring is allowed. The writes and reads are of data and/or parity blocks. Theprocessor20 includes areconstructor22 to reconstruct data blocks from faileddisks14a,14band14c. Thebus16 can have separate data and control lines (not shown) for each of thedisks14a,14band14c. The reads and writes may be parallel accesses to all or to a subset of thedisks14a,14band14c. In other embodiments a single set of data and control lines connects to eachdisk14a,14band14cof theRAID configuration12, and theprocessor20 performs serial writes to and reads from theseparate disks14a,14band14cover the shared data line. In this case, thebus16 may be a single SCSI bus or another type of shared or dedicated interconnect.
• The[0021]processor20 also includes areconstructor22. In some embodiments, theprocessor20 is programmed to provide thereconstructor22. In other embodiments, thereconstructor22 is implemented as a coprocessor, as discussed below.
• The[0022]processor20 transmits data blocks over an interface orline17, for example, a bus or a cable, airwaves, and so forth to a requestingdevice19. Theprocessor20 executes aprocess40 to produce a hint for use by theprocessor20 or in some embodiments, a remote receiving device (e.g., receiving device19) to indicate that a request for data may be directed to a disk that is bad or will have a slow response. The file system (described below) executes onprocessor20 has knowledge of how data is stored on disk drives. The file system has a mapping of data blocks to disk storage locations. The disk controller (not shown) returns a message to the file system indicating the state of the disk drives. With the knowledge of the mapping and message pertaining to the state of the disk drives, the file system can produce the hint to theprocessor20 or a requestingsystem19, as the case may be. Theprocessor20 or requestingsystem19 determines whether to read the block from the drive or initiate a recovery mode and read the corresponding parity block and data from other drives, thus avoiding the bad drive.Software components70 including the file system are described in FIG. 6.
• Referring to FIG. 2, a[0023]clustered system10 includes a plurality ofnodes11,11′,11″ includingrespective RAID configurations12,12′,12″ as in FIG. 1. The clusteredsystem10 is an architecture that stores data (e.g., for a video on demand embodiment, MPEG video and audio) across a set of loosely coupledcomputer systems11. Each of thenodes11,11′ and11″ include respectively ones of theRAID configurations12,12′ and12″, and local and remote file systems (discussed below) so that any one of the nodes can access data on any and/or all of the other nodes12-12″ in the clusteredsystem10. Thenodes11,11′,11″ are interconnected by anetwork35 so that any one of the cluster members can communicate directly with some or all of the other members of thecluster10. Exemplary networks include a point-to-point, or switched, packet, or cell networks, bus configurations, hyper-cube arrangements and so forth. In some embodiments, the disk drives themselves need not be a redundant RAID configuration or a RAID configuration. Exemplary redundant RAID configurations include RAID-5 and RAID-4.
• A first processor, e.g.,[0024]processor20 receives consecutive groups of pairs of data blocks and generates a parity block to associate with each pair of data blocks. Thefirst processor20 sends one block from each associated group of three blocks to each ofRAID configurations12,12′,12″. Eachprocessor20,20′,20″ of therespective RAID configuration12,12′,12″ subsequently breaks each block into two mini-blocks and generates a parity mini-block to associate with the two mini-blocks. Each secondlevel RAID configuration12,12′,12″ stores the mini-blocks as illustrated in FIGS. 3 and 4 onrespective disk drives14a-14c,14a′-14c′ and14a″-14c″. Thefirst processor20 can also retrieve blocks from theRAID configurations12,12′,12″ and deliver the retrieved blocks over an interface to a receivingdevice19.
• Still referring to FIG. 2, the clustered[0025]system10 handles slowly responding or missing or degraded devices by reconstructing and transmitting reconstructed blocks. Areconstructor22,22′,22″ respectively reconstructs and transmits to a receivingdevice19, the reconstructed block for any one of thedegraded RAID configurations12,12′,12″. More specifically, the reconstructor, e.g. reconstructor22 forprocessor20, reconstructs and transmits data for a degraded or slowly responding disk drive in one of thedegraded RAID configurations12,12′,12″. Adegraded RAID configuration12,12′,12″ is signaled by a hint sent to a requesting processor, e.g., thefirst processor20 of this example, from adegraded RAID configuration12,12′,12″. Thefirst processor20 andreconstructor22 handles latencies due to degraded disk responses. Thefirst processor20 provides predictability into the read latencies of theRAID configuration10. The receivingdevice19 can be embedded in eachconfiguration12 or can be an external device, e.g., a set top box for a video on demand application.
• The[0026]cluster10 uses a RAID process to build a fault tolerant clustered system of single processors. Data objects are stored as named fragment files across the members of the cluster. Data objects are striped in accordance with RAID-5 process (or other process, e.g., RAID-4) and stored with parity information to allow a missing named fragment file to be reconstructed if a cluster member (node)11 fails.
• Referring to FIG. 3, a[0027]fragment50 of a data stream from thesystem10 is shown. Thefragment50 will be stored in the RAID configurations12-12″ of FIG. 1. Theprocessor20 divides thefragment50 into an ordered sequence of blocks D(0), D(1), . . . D(i) and determines a parity block P(i, i+1) (i=0, 2, 4, . . . ) to associate withconsecutive pairs52,54 of the data blocks D(i), D(i+1). The parity block P(i, i+1) encodes at least one parity bit for each pair of equivalent bits of the associatedpair52,54 of data blocks D(i), D(i+1). The processor15 may write each associatedpair52,54 of data blocks D(i), D(i+1) and parity block P(i, i+1) to the three the RAID configurations12-12″ in parallel or serially as explained in FIG. 1.
• Referring to FIG. 4 the[0028]processor20 writes data and parity blocks to storage locations on the threedisks14a-14c,14a′-14c′ and14a″-14c″ in the RAID configurations12-12″, as shown. The data are arranged in stripes S1-S6. Each stripe S1-S6 stores a group of three associated blocks, which includes a consecutive pair of data blocks D(i), D(i+1) and the parity block P(i, i+1) constructed from the pair. In some RAID implementations, the portion of each configuration12-12″ in a particular stripe S1-S6 stores either one of the data blocks D(i), D(i+1) or the associated parity block P(i, i+1). Theprocessor20 writes the parity blocks P(i, i+1) associated with sequential pairs to different ones of the configurations12-12″ by cyclically shifting the storage location of P(i, i+1) in each consecutive stripe. This is referred to as rotating the parity blocks P(i, i+1) across the configurations12-12″. Rotating the storage location of the parity block more uniformly distributes the data blocks D(j) among the configurations12-12″ spreading access burdens more uniformly across the different configurations12-12″ during data reads and writes.
• The configuration shown in FIGS. 1 and 4 is often referred to as a RAID-[0029]5 configuration. Other configurations could be used such as RAID-4 and so forth. For example, in RAID-4 parity is stored on a dedicated parity drive of the RAID array, rather than dispersed among drives of the array, as in RAID-5. Clearly, other redundancy methods can be used including various forms of, for example, Huffman coding and other redundant coding techniques so that not only may one failure be taken into account but multiple failures can be taken into account. There may be an associated cost in increased processing in writing and reading of data.
• Referring to FIG. 5, the[0030]reconstructor22 includes amemory device62 and ahardware processor64. Both thememory device62 and theprocessor64 are coupled to a bus (not shown) that is part ofprocessor20. In some embodiments, theprocessor64 can be the same processor as20,20′ or20″ or can be a different processor, e.g., a coprocessor. Thememory device62 among other things receives data and/or parity blocks from the configurations12-12″ via thebus16. Thememory device62 stores the associated data and parity blocks for reconstructing the associated block of a slowly responding configuration12-12″.
• The[0031]processor64 performs an exclusive OR (XOR) of the associated parity and data blocks to reconstruct the data block of the stalled one of configurations12-12″. To perform the XOR, theprocessor64 reads the associated blocks from thememory device62. Theprocessor64 XOR's corresponding bits of the read associated parity and data blocks in a byte-by-byte manner. Theprocessor64 writes the results of the XOR back to thememory device62 for transmission to the appropriate one of the receiving devices39-39″. Thereconstructor22 can produce a reconstructed block for any one of the configurations12-12″.
• Referring now to FIG. 6,[0032]software components70 of theprocessors systems20,20′ and20″ are shown. Thesoftware components70 include aport driver71 and aclass driver72. In addition, the software includes file system including aremote file provider74, alocal file provider76, and a file system78 (SeaFile, FAT, NTFS). The software includes a SCSI driver80 and aRAID controller82, as well as afast Ethernet adapter84 and a SeaNet transport layer86.
• In order to provide transparent data object access, a RAID port driver masks that the data object exist as a set of named fragment files. The port driver provides multiplexing and demultiplexing services to merge named fragment files into a data stream. The RAID port driver registers as both a provider and a consumer. When the class driver attempts to open a data object, it calls all of the provider port drivers in the system. Upon being called, the RAID port driver becomes a consumer and uses the class driver to open each of the data fragments that comprise the data object. Once a session has been established to each of the named fragment files, that is, once access has been provided to each named fragment file, the RAID port driver performs an open call back to notify the class driver that the data object is available.[0033]
• In the particular application of accessing the data (e.g., video) objects, the port driver accesses the data (video) objects stored on the cluster. The remote file provider represents any third party application or device driver that might use the cluster technology. The on-disk structure of the cluster volume (file system[0034]54) can be either NTFS, FAT, SeaFile or raw disk access. The file system component is responsible for storing and retrieving the named fragment files.
• Referring to FIG. 7, a process[0035]100 to read data from thecluster10 is shown. Theprocessor20 selects102 the strip having the associated data blocks to transmit and supplies read-ahead information. The processor15 determines106 whether a hint from a previous access had indicated that any of the disks31-33 were faulty. If not, theprocessor20 delivers108 the requested data blocks. If the hint indicated that a drive31-33 was faulty, the reconstructor22 requests data block of the not faulty one of drives31-33 and parity to reconstruct the data block of the faulty disk drive31-33 without bothering to access the faulty drive. Thereconstructor22 receives the associated data and parity blocks from storage locations of the same stripe S1 of the other disks31-33, which are not slowly responding or degraded. Thereconstructor22 reconstructs the data and sends112 the reconstructed data block to the data requester19. Theprocessor20 saves114 hints returned from each node or RAID configuration12-12″ for future requests.
• The processor[0036]20 (FIG. 1) processes the request and at the same time supplies a read ahead message to the file system, so that the file system can read the next block and have it in a cache (not shown). The read ahead message indicates that the next time, the requester will read the offset+64 Kbytes, for example. Depending on the configuration, sometimes a 64 KByte (KB) block of data is skipped, e.g., for a RAID-5 implementation to skip parity data. On the other hand, in a RAID-4 implementation it is not necessary to skip blocks, since parity is on a separate drive and a request is made to the parity drive.
• The read-ahead message is used to show where the parity is. When the previous I/O transaction has completed, the file system on the[0037]processor20 has communicated with the RAID controller (not shown) and has determined if the next block would be valid (e.g., available from the RAID controller, will respond in time, and so forth) or not. The file system on theprocessor20 indicates via, the hint message, if the next block will be valid or invalid. If invalid, the file system returns the hint to the requestingprocessor20,20′20″ indicating that the next block will be a delayed read or will not be available, etc. This will enable the requesting one of configurations12-12″ to initiate a read around the bad block, i.e., to forgo a subsequent read request to the bad drive and rather to request parity and data from good drives.
• The hint message can indicate a “next bad byte offset.” The next bad byte offset indicates where the bad block is located based on some offset. The requesting system receives the “next bad byte offset” and can determine that the RAID array has a bad or reconstructing drive. The requesting system will skip the drive and instead read the parity and data block from the other drives. The[0038]reconstructor22 performs an XOR operation on parity and data to reconstruct the missing data corresponding to the bad or reconstructing drive. Thus, by skipping the bad drive and not waiting for a slow drive to respond before trying to reconstruct data, will improve the performance of the RAID array towards normal performance.
• Returning of the hint instructs the requesting system to start the recovery process or requesting process without waiting for degradation of performance to occur. The[0039]processor20 through the file system has intimate knowledge of the block mapping and file structure on theRAID configuration11 and thedisk drives14a-14c. Thus, theprocessor20 can examine the request address and a stored block map (not shown) that indicates where the blocks are stored on the disk drives31-33. From that information the file system constructs the hint to send to a device that will request the bad block to indicate the state of the RAID array. The hint masks various failures of the RAID array.
• [0040]Nodes11,11′ and11″ in thecluster10 are peer-to-peer, consuming and providing data to the clusteredsystem10. The hint improves the performance of an impaired disk subsystem. When a drive is dead or rebuilding in a RAID array, theprocessor20 needs to read the entire stripe row to reconstruct the data missing on the failed drive. When a large number of reader threads is placed on a degraded RAID array, the reconstruction cache is purged as new requests are serviced. Because the RAID array has a fixed size reconstruction cache, some/all data blocks may need to be read twice because they are flushed before they are used to reconstruct the bad block. The performance of the RAID controller will degrade to ½ or more of the non-degraded performance in this case. By using the hint message to avoid the failed drive, the impaired RAID array can run at {fraction (11/12)}th efficiency (assuming a 12 drive array). The input/output requests from the failed drive are handled by the other systems in the cluster in the form of parity data requests ({fraction (1/12)}th of the load on the RAID array).
• The process[0041]100 reconstructs the block from parity and data as discussed above. With the hint indicating a bad read, the read ahead is turned off, because performing such a read operation will degradesystem10 performance. The read ahead cache (not shown) is disabled at the point that the file system returns the hint to the requesting system indicating the next request at the byte offset will land on a bad drive.
• In other embodiments the[0042]file system74 can look further ahead and produce a hint to theremote system19 where the next request to the bad drive is going to land, so that the requestingsystem11 can read up to that point and skip over the bad drive. Thelocal file system74 on theprocessor20 manages the file fragment at the remote end and knows that the next block in the file is either bad or good. Thefile system74 returns the hint to the requestingsystem11. The requestingsystem19 determines whether to read the block from the drive or initiate a recovery mode and read the parity block and data from other drives and Exclusive-Or the data with parity to recover the lost block.
• In a case where the[0043]file system74 encounters the same situation with two systems, thefile system74 can implement a tie-breaking algorithm, e.g., a round robin algorithm where thesystem10 first reads one node and the next time reads from the other node.
• Referring back to FIGS.[0044]1-3, theprocessor20 controls reconstruction and transmission of reconstructed data blocks. Theprocessor20 orders the remaining nodes12-12″ to transmit the associated blocks to thereconstructor22, e.g., to thememory device62, if the received hint indicates a problem. In FIG. 4, the associated data and parity blocks are stored in the same stripe S1-S6 as the untransmitted data block from the defective or slowly respondingdisk14a-14c. Thus, theprocessor20 orders reads of the associated stripe S1-S6 to obtain the associated blocks. Theprocessor20 signals thereconstructor22 to reconstruct the data block from a slowly responding disk, e.g., by a signal sent to theprocessor64 of FIG. 3. Then, theprocessor20 reads the reconstructed block from thereconstructor22, e.g., thememory device62, and transmits the reconstructed block via the interface orline17.
• The[0045]processor20 does not interrupt a slowly respondingdisk14a-14cfrom recovering by sending a request to transmit data. Instead, theprocessor20 orders thereconstructor22 to reconstruct the missing data from the associated data blocks in the normally responding nodes12-12″.
• Referring to FIG. 8, a[0046]video transmission system150, which uses thesystem10 of FIG. 1 is shown. Areceiver152 receives data blocks transmitted from the interface vialine17 at aninput terminal156. Transmission between thesystem10 andreceiver152 may be by radio wave, light, and/or cable transmission. Theinput terminal156 couples to aninput data buffer154, e.g., a first-in-first-out buffer. Theinput data buffer154 stores two to several times the quantity of data included in one data block shown in FIG. 2B. Data stored in theinput data buffer154 provides for continuous video data processing in the event of a short transmission interruption.
• The[0047]video transmission system150 can lower the occurrence of viewing pauses by transmitting a reconstructed data block in response to receiving the hint. In one embodiment of thesystem150, theRAID configuration11 insystem10 needs about 100 ms to transmit and reconstruct a data block. The receiver's input data buffer117 stores about 2000 ms of video data. This storage provides sufficient stored data to deliver normal video while giving sufficient time to reconstruct lost data in response to receiving the hint.
• Other embodiments are within the scope of the appended claims.[0048]

Claims (17)

What is claimed is:

1. A method for reading data from a clustered system of a plurality of redundant array of inexpensive disks (RAID) device configurations, the method comprises:

receiving a message from a requesting one of members of the clustered system that indicates a next location for a particular block for a subsequent read operation from a drive in the RAID device; and

returning to the requesting one of members of the clustered system a hint that indicates the state of the drive in the RAID device corresponding to location for the subsequent read operation.

2. The method ofclaim 1 wherein the hint indicates that performance of the device is degraded or the device is unavailable.

3. The method ofclaim 1 wherein the system returns a next bad byte offset message to the requesting device.

4. The method ofclaim 1 wherein the requesting member will skip the drive and instead read redundant information and at least one data block one of the other drives in the RAID device.

5. The method ofclaim 1 wherein the requesting member will reconstruct data from the skipped drive from read redundant information from one of the other drives in the RAID device.

6. The method ofclaim 1 further comprising:

performing an XOR operation on parity and data to reconstruct the missing data corresponding to data on the skipped drive.

7. The method ofclaim 1 further comprising:

requesting a first disk of the RAID configuration to transmit a first data block stored therein to a receiving device;

reconstructing the first data block from associated data stored in other disks of the RAID configuration; and

transmitting the reconstructed first data block directly to the receiving device.

8. The method ofclaim 7 wherein the reconstructing and the transmitting are performed in response to receiving the hint indicating that performance of the device is degraded or the device is unavailable.

9. The method ofclaim 6 wherein a disk returns a next bad byte offset, and the requesting system receives the next bad byte offset and determines that the RAID array has a bad or reconstructing drive.

10. The method ofclaim 1 wherein returning the hint instructs the requesting member to start a recovery process or requesting process without waiting for degradation of performance to occur.

11. A clustered system of a plurality of redundant array of inexpensive disks (RAID) configurations to transmit data blocks to a receiving device, comprising:

a plurality of cluster members coupled so that any one cluster member can deliver data to at least some of remaining cluster members, each cluster member comprising:

a plurality of disks adapted to store the blocks and to transmit the stored blocks to the receiving device;

a processor to control reads from and writes to the plurality of disks; and

a process to receive a message that indicates a next location for a particular block for a subsequent read operation from a drive in the RAID device and produce a hint that indicates the state of the drive in the RAID device corresponding to location for the subsequent read operation.

12. The configuration ofclaim 11 further comprising:

a device to reconstruct a particular block from associated blocks and to transmit the reconstructed block to a receiving device in response to the hint that indicates the state of the drive in the redundant arrays of inexpensive disks.

13. The configuration ofclaim 11 wherein the processor starts a recovery process based on a state of the hint.

14. The configuration ofclaim 11 wherein a disk controller returns a next bad byte offset, and the requesting system receives the next bad byte offset and determines that the RAID array has a bad or reconstructing drive.

15. A clustered configuration comprises:

a plurality of clustered nodes each node comprising a redundant array of inexpensive disks (RAID) and a processor adapted to write blocks to and read blocks from the arrays; and

a process to receive a message that indicates a next location for a particular block for a subsequent read operation from a drive in the redundant arrays of inexpensive disks of one of the nodes and produce a hint that indicates the state of the drive in the redundant arrays of inexpensive disks corresponding to location for the subsequent read operation.

16. The configuration ofclaim 15 further comprising:

a device to reconstruct a particular block from associated blocks and parity information and to transmit the reconstructed block to a receiving device in response to the hint that indicates the state of the drive in the second level redundant arrays of inexpensive disks.

17. The configuration ofclaim 15 wherein the hint indicates that performance of the device is degraded or the device is unavailable.

Images